Estimated reading time: 7 minutes · Last updated:
Suspected China-linked attackers pointed autonomous AI agents at MONAI Label, an open-source server that imaging researchers rely on to annotate scans. The agents found a previously unknown path-traversal bug and used it to reach a root shell on a university machine, Amazon Threat Intelligence said in November 2026. The MONAI Label zero-day, now catalogued as CVE-2026-133968 with a critical 9.8 score, lets anyone who can reach an unauthenticated server write files beyond its storage directory. Versions 0.8.5 and earlier are exposed in their default configuration, where login is switched off. Maintainers merged a source-code correction on 19 October, yet no installable release carries it.
Key takeaways
- Root on one university server: Amazon Threat Intelligence says AI agents run through the Cairn framework found a MONAI Label path-traversal zero-day and turned it into root access on a single university server.
- Vulnerable out of the box: MONAI Label 0.8.5 and earlier ship with authentication disabled, so any client that can reach the server over the network can write files outside its storage folder.
- Minutes, not weeks: Across about 48 hours in September, the agents mapped the application and got to a working file write in roughly 11 minutes, with the operator supplying only the goal and the targets.
- Critical score, nothing to install: CVE-2026-133968 is rated 9.8 out of 10, and a fix merged on 19 October, but PyPI still offers only version 0.8.5, released in 2024.
- Attribution held at moderate: Amazon cites Simplified Chinese prompts and China Standard Time settings in linking the work to a China-linked actor, but names no group and does not claim state sponsorship.
Table of contents
- Key takeaways
- How the MONAI Label Path-Traversal Flaw Lets Outsiders Write Files
- What the Attackers' Exposed Cairn Panel Revealed to Amazon
- How an 11-Minute File Write Became a Root Shell
- Why a Merged Fix Is Not Yet a Patch
- Protecting MONAI Label Servers Before a Release Ships
- What helps and what hurts containment
- What to be careful about
- Frequently asked questions
How the MONAI Label Path-Traversal Flaw Lets Outsiders Write Files
MONAI Label is a standalone server, separate from the wider MONAI framework, that research groups run to annotate medical scans and train imaging models. The bug lives in the code that saves uploaded images and labels. A path-traversal flaw is one where a filename supplied by a user can climb out of the folder a program meant to confine it to.
Here, the server takes a file identifier from the caller, joins it to its storage directory and saves the result without checking that the final path still sits inside that directory. Because authentication ships disabled, any client with network reach can trigger the write. Versions 0.8.5 and earlier are affected.
The damage has a ceiling. Files land with whatever permissions the server's own account holds, so an attacker cannot place files anywhere they please. If that account can write to a folder holding code or configuration, though, a file write becomes a route to running code.
What the Attackers' Exposed Cairn Panel Revealed to Amazon
Amazon rebuilt the intrusion from an unusual vantage point: the operator's own Cairn dashboard, which sat on the open internet with no password protecting it. Cairn, an open-source tool, hands the work of penetration testing to AI agents and automates it. Through the exposed panel, Amazon could see how the agents were configured, which prompts they were given and a log of every run.
Over roughly 48 hours in September, the agents charted the application's interface, and Amazon said they arrived at the file-write stage in about 11 minutes. The human operator chose the targets and set shell access as the objective. The agents picked their own techniques with little moment-to-moment steering, the company said.
Amazon's confidence that a China-linked actor ran the operation is moderate. It points to prompts written in Simplified Chinese, clocks set to China Standard Time and targets China would plausibly care about. It has tied the activity to no named group and has not established that a government directed it.
How an 11-Minute File Write Became a Root Shell
Root access did not come inside those 11 minutes. In a later, separate session, the agents reused the traversal bug to place a file that Python executes automatically whenever it starts. They then launched one of MONAI Label's training jobs, which spawned a fresh Python process. That process loaded the planted file and called back to the attacker with root privileges.
That outcome was recorded on one university server. Whether another installation would hand over the same access depends on how it is deployed and what rights its service account holds. The campaign also included operations that failed, so the root result is not a forecast for every deployment.
Later, during another session, the operator opened a tunnel for the agents into a second university's network, one that ran identical software. Amazon suspects the attackers broke into it using the same technique, though it has not confirmed that compromise.
Why a Merged Fix Is Not Yet a Patch
The MONAI Label maintainers acknowledged the bug and merged a correction, pull request 1915, on 19 October. The change forces every upload to resolve to a location inside the permitted folders and refuses filenames that are empty or begin with a dot.
Shipping it is another matter. The project's GitHub security advisory, GHSA-qc22-rmgc-vg94, went up in November 2026 and names no patched version, saying the correction will arrive in a later release. PyPI still lists 0.8.5, a 2024 release, as the current package. Code merged into a repository does nothing for teams that install from the package index.
The CVE record now carries the flaw as CVE-2026-133968, rated 9.8 out of 10, which falls in the critical band. The project's advisory has not caught up with that identifier yet; the mismatch is slow bookkeeping, not evidence of a second vulnerability.
Protecting MONAI Label Servers Before a Release Ships
With nothing to install, defence comes down to shrinking exposure and limiting what a hijacked server could touch. Network controls come first: firewall and routing rules that keep untrusted clients away from MONAI Label, rather than faith in URLs that are merely hard to guess.
Switching on authentication closes the door the default configuration leaves open. Running the server under a low-privilege account instead of root, and narrowing the directories it may write to, breaks the exact chain Amazon described, in which a planted startup file inherited root.
Teams building from source should confirm their build contains the 19 October change before calling it fixed. Any server that was reachable from untrusted networks deserves a check for unfamiliar Python startup files and training jobs nobody scheduled, the two traces the reported attack would leave behind.
| Source | What it shows | Date |
|---|---|---|
| GitHub pull request 1915 | Fix merged; uploads confined to approved folders, empty or dot-prefixed names rejected | 19 October |
| GitHub advisory GHSA-qc22-rmgc-vg94 | No patched version; fix promised for a future release; no CVE listed | November 2026 |
| PyPI release listing | Current package is still 0.8.5 | 2024 |
| CVE record | CVE-2026-133968, rated 9.8 out of 10 | November 2026 |
What helps and what hurts containment
The case for
- The merged pull request already defines the correct behaviour, so a release mainly needs packaging rather than new engineering.
- Authentication, firewall rules and a non-root service account are available today and directly cut the chain Amazon reconstructed.
- The exposed Cairn panel gave Amazon the agents' prompts, task settings and run history, so defenders know the specific traces to hunt for.
The case against
- No installable version newer than 0.8.5 exists, and PyPI still serves the vulnerable 2024 package.
- Authentication is off by default, so any server deployed without changes is reachable by unauthenticated clients.
- The agents needed little direction from the operator, chose their own methods and reached a file write in about 11 minutes.
What to be careful about
- Research servers running MONAI Label 0.8.5 or earlier in the default no-login configuration can be written to by any client that reaches them over the network.
- A source build may be treated as patched when it predates the 19 October fix.
- Deployments whose service account runs as root or can write to Python startup locations are open to the same escalation path Amazon described.
- Amazon's account is its own reconstruction without independent confirmation, and the second university compromise remains unverified.
The bottom line
Amazon's account comes from its own reconstruction, and nobody has independently confirmed it. Nothing in it demonstrates that scans or patient records left the targeted universities, that clinical systems were reached, or that the attacker went past the university targets described. What it does show is narrower and still uncomfortable: an operator set a goal, picked targets and let Cairn agents find and chain a path-traversal bug in an unauthenticated research server, reaching a file write in about 11 minutes. The weak point was configuration as much as code, since MONAI Label ships with login off. Until maintainers cut a release past 0.8.5, protection rests on the network, the service account and the audit.
What to watch
- Watch for a MONAI Label release newer than 0.8.5 that ships the 19 October fix; the maintainers say it will come in a future release but have set no date.
- Watch for Amazon to confirm or rule out the suspected compromise of the second university network; no timeline has been given.
- Watch for the GHSA-qc22-rmgc-vg94 advisory to be updated with CVE-2026-133968 and a patched version number; no date has been set.
Frequently asked questions
Is MONAI Label the same as the MONAI framework?
No. MONAI Label is a separate server program used to label medical scans and train imaging models. CVE-2026-133968 concerns MONAI Label 0.8.5 and earlier, not the broader MONAI framework.
Were patient records stolen in the MONAI Label attack?
Amazon's reconstruction does not show that scans or patient records were taken or that clinical systems were touched. What it describes is root access on one university server and a suspected, unconfirmed breach of a second university network.
Is there a patch for CVE-2026-133968?
A source fix, pull request 1915, merged on 19 October, but advisory GHSA-qc22-rmgc-vg94 lists no patched version. PyPI still offers 0.8.5, so teams must rely on authentication, network blocking and a non-root account until a release appears.
What is Cairn and how did it expose the attackers?
Cairn's purpose is to automate penetration testing by delegating the job to AI agents, and its code is open source. Because the operator never put a password on a Cairn control panel exposed online, Amazon was able to read the agents' task settings, their prompts and their run history.
Related reading